Abstract
When isolated word speech recognition devices are deployed in real user applications, they must do more than simply recognize isolated words. Real users often embed target words in a longer utterance, and the words should be recognized as often as possible. Out-of-vocabulary utterances must be rejected. The target application is to recognize the city name query of directory assistance. Within this project, a keyword spotting algorithm has been developed which markedly improves the recognizer's performance on embedded speech. A rejection algorithm has also been implemented, and several variations on the algorithmic approach are compared with respect to their ability to reject out-of-vocabulary utterances while still accepting in-vocabulary correctly-recognized utterances. The keyword spotting and rejection algorithms are evaluated, and the factors contributing to their success are described.

This publication has 5 references indexed in Scilit: